Widespread generative AI adoption could raise annual US labor productivity growth by about 1.5 percentage points over ten years
Assessment
Credible evidence or argument exists on multiple sides.
The figure originates with Goldman Sachs economists Briggs and Kodnani (March 2023), who project that widespread generative AI adoption could lift annual US labor productivity growth by about 1.5 percentage points over a decade; Goldman reaffirmed the estimate in 2025 as a roughly 15 percent lift in the productivity level at full adoption. The projection rests on a task-exposure calculation: the estimate that generative AI exposes the equivalent of some 300 million full-time jobs to automation is broadly corroborated across independent studies, and field evidence of 14-15 percent productivity gains for AI-assisted customer support agents shows task-level gains of the assumed size are achievable. But exposure measures what could be automated, not what will be, and the projection's critical steps, that a large share of exposed work is profitably automated and adopted within ten years, are exactly where credible economists diverge.
The principal rival estimate, Acemoglu's bound of under one percent total factor productivity gain over ten years, reaches a figure an order of magnitude lower from similar exposure data by assuming only a small fraction of exposed tasks is economically automatable in the window; that bound is itself contested, with its low-end inputs drawing substantial criticism. The historical record of general-purpose technologies taking decades to show aggregate gains and a Danish study finding no detectable effect of chatbot adoption on earnings or hours within two years weigh against the projected pace, though the latter measures compensation rather than output per hour.
Direct evidence remains inconclusive as of mid-2026. US productivity growth has accelerated since 2022, but whether generative AI is a significant contributor to that acceleration is genuinely disputed: utilization-adjusted TFP has been nearly flat, and Goldman's own 2026 analysis finds no meaningful economy-wide relationship between AI and productivity yet, while newer regional Federal Reserve analyses find higher-adoption industries growing faster. Sustained productivity growth near or above 2.5 percent through the late 2020s with AI-exposed industries visibly leading would move the claim toward supported; continued absence of a differential signal as the window closes would move it toward contradicted.
Full reasoning: the evidence and decisions behind this verdict
The verdict remains contested after re-judging against three newly assessed subclaims; none forces a status change, and they pull in partially offsetting directions.
First, the 300-million-jobs exposure estimate is now assessed supported (0.8), replacing its previously unassessed standing. This firms up the projection's input layer but with an explicit caveat that matters here: exposure is not predicted automation, so the supported exposure figure constrains this projection only through the further assumption that exposed work is actually automated within the window. That assumption, not the exposure data, is where the Goldman-Acemoglu gap lives, so the for-argument is no stronger where it was weakest.
Second, Acemoglu's under-one-percent TFP bound is now assessed contested with credence 0.30 that the low bound is right, its load-bearing inputs (task share under five percent, roughly 25 percent average cost savings, no large new-task gains) each drawing credible criticism. This mildly strengthens the high-end side: the strongest single counter-estimate is itself judged more likely wrong than right. But a 0.30 credence on the low bound is far from vindicating 1.5pp; the space between 0.66 percent level and 15 percent level is wide, and most intermediate estimates (e.g., in the 1-4 percent level range over a decade) would still leave this claim's figure too high.
Third, the post-2022 acceleration attribution subclaim is assessed contested with credence 0.35. Three-plus years into the ten-year window, no large AI effect is visible in aggregate data (utilization-adjusted TFP nearly flat through 2026Q1), which weighs against the projected pace; but early cross-industry evidence from Kansas City and Dallas Fed analyses shows higher-adoption industries growing faster, which is what the front edge of a J-curve would look like. This is genuinely two-sided and consistent with either late-arriving gains or a much smaller effect.
A recency check (March 2026 reporting on Goldman's own analysis, finance.yahoo.com/news/goldman-finds-no-meaningful-relationship-143553714.html) finds Goldman acknowledging no meaningful economy-wide AI-productivity relationship yet while maintaining the forward projection; this is about currently visible effects, not an assertion or retraction of the ten-year claim, so it is context rather than a new instance on either side.
Instance stances are unchanged and remain split: Goldman affirming twice (2023, 2025), Acemoglu denying (NBER WP 32487). Credible instances on both sides, rival estimates an order of magnitude apart on defensible assumption differences, and early data that underdetermine the answer: contested, with confidence nudged to 0.87 because the newly assessed subclaims confirm the shape of the dispute rather than complicating it. No claim credence is given: the modal "could," the conditional adoption timeline, and the unresolved attribution question make a single number false precision. Directionally, the burden still sits with the high-end projection, since independent academic estimates cluster well below 1.5pp and no aggregate signal of the required size has appeared; but the leading low-end rival is itself judged more likely wrong than right, so the honest summary is a wide distribution over intermediate outcomes.
What would change the verdict: sustained US productivity growth near or above 2.5 percent with AI-exposed industries visibly leading (toward supported); a closing window with no differential signal, or authorial downward revision (toward contradicted).
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because generative AI could expose the equivalent of some 300 million full-time jobs to automation, amounting to roughly a quarter of work tasks in advanced economies, and given that field evidence shows a generative AI assistant raising customer support agents' productivity by roughly 14-15%, applying cost savings of that order across the exposed share of work implies an aggregate lift in annual US labor productivity growth of about 1.5 percentage points once adoption is widespread.
The arithmetic goes through only under two unstated assumptions: that a large share of exposed tasks is profitably automatable within a decade, and that adoption becomes widespread on that timeline; both are exactly where critics diverge. The 300-million-jobs exposure estimate is now assessed as supported, but it measures exposure rather than predicted automation, so firming it up does not close the argument's real gap. The customer support experiment shows the assumed task-level gains are achievable but comes from a single setting and cannot by itself carry the generalization to the whole exposed task base.
Because independent task-based estimates bound AI's total factor productivity gain below one percent over ten years, because general-purpose technologies historically took decades to produce measurable aggregate productivity gains, and because two years of chatbot adoption produced no detectable effect on adopting workers' earnings or hours, the projected lift is an order of magnitude too large for its ten-year window.
Granting its premises, the argument makes the projected magnitude and pace unlikely rather than impossible: a lift could still arrive late in the window. Its weight rests chiefly on the estimate that AI raises total factor productivity by less than one percent over ten years, which is now assessed as contested with the balance of criticism running against its low-end inputs, so the argument's strongest premise carries less force than its order-of-magnitude framing suggests. The historical record of slow diffusion for general-purpose technologies supplies the more durable weight on pace, and the Danish null on earnings and hours is well supported but bears on the claim only as a proxy, since compensation can sit still while output per hour moves.
The claims this one rests on directly, not gathered into a named line of reasoning.
- supportsthis provides evidence for the parentsteward instructions →Generative AI is a significant contributor to the post-2022 acceleration in US labor productivity growth ↗︎ · shared subclaim
Provenance
Where this claim has been said, linked to its canonical form.
lift productivity growth by 1.5 percentage points over a 10-year period
they could drive a 7% (or almost $7 trillion) increase in global GDP and lift productivity growth by 1.5 percentage points over a 10-year period.
Using existing estimates on exposure to AI and productivity improvements at the task level, these macroeconomic effects appear nontrivial but modest—no more than a 0.66% increase in total factor productivity (TFP) over 10 years.
Acemoglu's task-based model, explicitly contrasted with the Goldman Sachs and McKinsey projections, argues the ten-year aggregate productivity effect is an order of magnitude below a 1.5pp annual lift.
Our economists estimate that generative AI will raise the level of labor productivity in the US and other developed markets by around 15% when fully adopted and incorporated into regular production.
Goldman Sachs Research reaffirming its projection of generative AI's labor-market and productivity effects; a 15% lift in the productivity level when fully adopted restates the 1.5pp-per-year-over-a-decade figure.
Assessment history
0 status changes over 2 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Aug 10, 2026. Every judgment on this page is accompanied by a reasoning trace.